An alternative approach for the classification of orange varieties based on near infrared spectroscopy

Abstract

A multivariate technique and feasibility of using near infrared spectroscopy (NIRS) for non-destructive discriminating Thai orange varieties were studied in this paper. Short-wavelength near infrared (SW-NIR) spectra in region of 643 to 970nm were collected from 100 orange sample of each varieties. A total of 300 spectra were used to develop an accurate classification model by diversity of classifiers. The result showed that Logistic Regression (LGR) model was achieved 100% classification accuracy while Multi-Criteria Quadratic Programming (MCQP) and Support Vector Machine (SVM) ones also demonstrated satisfying result (95%). In order to find simpler and easier interpretable classification model, several feature selection techniques were evaluated to identify the most relevant wavelengths to the orange varieties. With four principal components (PCs) from Principal Component Analysis (PCA) and the effective wavelengths of 769.68, 692.28, 662.61 and 959.31nm from Least Square Forward Selection (LS-FS), the reduced classification models of LGR also achieved satisfying classification accuracy respectively. Furthermore, both Kernel Principal Component Analysis (KPCA) and Kernel Least Square Forward Selection (KLS-FS) with SVM enhanced performance of models by 5 PCs and features respectively. The result concluded that NIRS can yield an accurate classification for Thai tangerine varieties from whole spectra and can enhance interpretability of classification model by feature subset.